import shutil
import sys
sys.path.insert(0, './src')
from models.classifier import FoodClassifier
from project_utils import calculate_metrics
import plot

# Initialize the classification model
model = FoodClassifier()

# Model testing, predict a single image
example_image = './data/example.JPG'
results = model.predict_single(example_image)
plot.single_image(example_image, title=f"{results['label']}: {results['confidence']:.2f}")

Model initialized, using device: cuda

dataset_path = './data/raw'
results = model.predict_folder(dataset_path)
plot.wrong_predictions(results, folder = dataset_path)
metrics = calculate_metrics(**results)
print(metrics)

Loaded 205 images from ./data/raw
Predicted 205 images in 8.03s

{'accuracy': 0.5658536585365853, 'f1_score': 0.6658446700550718}

from preprocessing.lowlight import gamma, CLAHE, SSRetinex
from preprocessing.deblurr import ssk, usm, swf
from preprocessing.downscaling import Lanczos, Lanczos_SAID, DPID
from project_utils import preprocess_folder

processor = model.processor
processor.do_resize = True
processor.do_center_crop = False

output_path = './data/preprocessed'

# Low-light
for func in [gamma, CLAHE, SSRetinex]:
    out_dir = output_path + func.__name__
    preprocess_folder(func, input_dir=dataset_path, output_dir=out_dir)
    results = model.predict_folder(out_dir)
    metrics = calculate_metrics(**results)
    print(f"Metrics for {func.__name__}: {metrics}")
    plot.wrong_predictions(results, folder = out_dir)
# shutil.rmtree(output_path)

processor.do_resize = True
processor.do_center_crop = False

# #Deblurring
for func in [ssk, usm, swf]:
    out_dir = output_path + func.__name__
    preprocess_folder(func, input_dir=dataset_path, output_dir=out_dir)
    results = model.predict_folder(out_dir)
    metrics = calculate_metrics(**results)
    print(f"Metrics for {func.__name__}: {metrics}")
    plot.wrong_predictions(results, folder = out_dir)
# shutil.rmtree(output_path)

processor.do_resize = False
processor.do_center_crop = False

# Downscaling
for func in [Lanczos, Lanczos_SAID, DPID]:
    out_dir = output_path + func.__name__
    preprocess_folder(func, input_dir=dataset_path, output_dir=out_dir)
    results = model.predict_folder(out_dir)
    metrics = calculate_metrics(**results)
    print(f"Metrics for {func.__name__}: {metrics}")
    plot.wrong_predictions(results, folder = out_dir)
# shutil.rmtree(output_path)

Loaded 205 images from ./data/preprocessedgamma
Predicted 205 images in 16.12s
Metrics for gamma: {'accuracy': 0.5756097560975609, 'f1_score': 0.6678271677670021}

Loaded 205 images from ./data/preprocessedCLAHE
Predicted 205 images in 16.33s
Metrics for CLAHE: {'accuracy': 0.5756097560975609, 'f1_score': 0.6880086635284361}

Loaded 205 images from ./data/preprocessedSSRetinex
Predicted 205 images in 17.18s
Metrics for SSRetinex: {'accuracy': 0.2975609756097561, 'f1_score': 0.4352442714584815}

Loaded 205 images from ./data/preprocessedssk
Predicted 205 images in 16.88s
Metrics for ssk: {'accuracy': 0.551219512195122, 'f1_score': 0.6517573080750508}

Loaded 205 images from ./data/preprocessedusm
Predicted 205 images in 17.65s
Metrics for usm: {'accuracy': 0.551219512195122, 'f1_score': 0.6554786566545715}

Loaded 205 images from ./data/preprocessedswf
Predicted 205 images in 14.86s
Metrics for swf: {'accuracy': 0.45365853658536587, 'f1_score': 0.5325994826796637}

Loaded 205 images from ./data/preprocessedLanczos
Predicted 205 images in 6.62s
Metrics for Lanczos: {'accuracy': 0.5756097560975609, 'f1_score': 0.6702244717777451}

[INFO] Loading SAID model 'SAID_Lanczos' from c:\Users\Austin Chen\Desktop\UCSD\2025_Fall\ECE253\Project\ECE253-food-classification\src\preprocessing\..\..\SAID\pretrained_models\SAID_Lanczos.pth

c:\Users\Austin Chen\Desktop\UCSD\2025_Fall\ECE253\Project\ECE253-food-classification\./src\preprocessing\downscaling.py:74: FutureWarning: You are using `torch.load` with `weights_only=False` (the current default value), which uses the default pickle module implicitly. It is possible to construct malicious pickle data which will execute arbitrary code during unpickling (See https://github.com/pytorch/pytorch/blob/main/SECURITY.md#untrusted-models for more details). In a future release, the default value for `weights_only` will be flipped to `True`. This limits the functions that could be executed during unpickling. Arbitrary objects will no longer be allowed to be loaded via this mode unless they are explicitly allowlisted by the user via `torch.serialization.add_safe_globals`. We recommend you start setting `weights_only=True` for any use case where you don't have full control of the loaded file. Please open an issue on GitHub for any issues related to this experimental feature.
  ckpt = torch.load(ckpt_path, map_location="cpu")
c:\Users\Austin Chen\Desktop\UCSD\2025_Fall\ECE253\Project\ECE253-food-classification\./src\preprocessing\downscaling.py:226: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(enabled=amp_enabled, dtype=torch.float16):
c:\anaconda250618\envs\ece253_env\Lib\site-packages\torch\functional.py:534: UserWarning: torch.meshgrid: in an upcoming release, it will be required to pass the indexing argument. (Triggered internally at C:\cb\pytorch_1000000000000\work\aten\src\ATen\native\TensorShape.cpp:3596.)
  return _VF.meshgrid(tensors, **kwargs)  # type: ignore[attr-defined]

Loaded 205 images from ./data/preprocessedLanczos_SAID
Predicted 205 images in 6.92s
Metrics for Lanczos_SAID: {'accuracy': 0.5560975609756098, 'f1_score': 0.6504567606954238}

preprocess_folder(CLAHE, input_dir=dataset_path, output_dir=output_path)
preprocess_folder(ssk, input_dir=output_path, output_dir=output_path)
# model.processor.do_resize = False
# model.processor.do_center_crop = False
preprocess_folder(Lanczos, input_dir=output_path, output_dir=output_path)

model.processor.do_resize = False
model.processor.do_center_crop = False

results = model.predict_folder(output_path)
metrics = calculate_metrics(**results)
print(f"Metrics for best combination: {metrics}")
plot.wrong_predictions(results, folder = output_path)
shutil.rmtree(output_path)

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 6.60s
Metrics for best combination: {'accuracy': 0.4292682926829268, 'f1_score': 0.5669924198829516}

def Identity(image):
    return image
# find the best low-light - deblur - downscale combination

model.processor.do_resize = False
model.processor.do_center_crop = False

output_path = './data/preprocessed'
best_metrics = None
best_combination = None

for func1 in [Identity, gamma, CLAHE, SSRetinex]:
    for func2 in [Identity, ssk, usm, swf]:
        for func3 in [Lanczos, Lanczos_SAID, DPID]:
            preprocess_folder(func1, input_dir=dataset_path, output_dir=output_path)
            preprocess_folder(func2, input_dir=output_path, output_dir=output_path)
            preprocess_folder(func3, input_dir=output_path, output_dir=output_path)
            results = model.predict_folder(output_path)
            metrics = calculate_metrics(**results)
            print(f"Metrics for {func1.__name__} + {func2.__name__} + {func3.__name__}: {metrics}")
            plot.wrong_predictions(results, folder = output_path)
            shutil.rmtree(output_path)

            if best_metrics is None or metrics['accuracy'] > best_metrics['accuracy']:
                best_metrics = metrics
                best_combination = (func1.__name__, func2.__name__, func3.__name__)

print(f"Best combination: {best_combination} with metrics: {best_metrics}")

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 6.40s
Metrics for Identity + Identity + Lanczos: {'accuracy': 0.5756097560975609, 'f1_score': 0.6702244717777451}

c:\Users\Austin Chen\Desktop\UCSD\2025_Fall\ECE253\Project\ECE253-food-classification\./src\preprocessing\downscaling.py:226: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(enabled=amp_enabled, dtype=torch.float16):

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 7.15s
Metrics for Identity + Identity + Lanczos_SAID: {'accuracy': 0.5560975609756098, 'f1_score': 0.6504567606954238}

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 6.47s
Metrics for Identity + Identity + DPID: {'accuracy': 0.5024390243902439, 'f1_score': 0.6202426763231333}

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 6.38s
Metrics for Identity + ssk + Lanczos: {'accuracy': 0.5365853658536586, 'f1_score': 0.6369884345577189}

c:\Users\Austin Chen\Desktop\UCSD\2025_Fall\ECE253\Project\ECE253-food-classification\./src\preprocessing\downscaling.py:226: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(enabled=amp_enabled, dtype=torch.float16):

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 6.95s
Metrics for Identity + ssk + Lanczos_SAID: {'accuracy': 0.5463414634146342, 'f1_score': 0.6551917573494846}

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 6.46s
Metrics for Identity + ssk + DPID: {'accuracy': 0.32682926829268294, 'f1_score': 0.45701842377303065}

Loaded 205 images from ./data/preprocessed
Predicted 205 images in 6.26s
Metrics for Identity + usm + Lanczos: {'accuracy': 0.5463414634146342, 'f1_score': 0.6486484452115001}

c:\Users\Austin Chen\Desktop\UCSD\2025_Fall\ECE253\Project\ECE253-food-classification\./src\preprocessing\downscaling.py:226: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(enabled=amp_enabled, dtype=torch.float16):

model.processor.do_resize = True
# model.processor.do_center_crop = True  


from models.trainer import fine_tune

fine_tune(
    model_wrapper=model,
    data_dir=dataset_path,
    val_ratio=0.3,
    epochs=20, 
    batch_size=4,
    lr=5e-5
)

Loaded 205 images from ./data/raw
Epoch 1/20 - Loss: 1.5862 - Accuracy: 0.6643
Validation accuracy: 0.6774
Epoch 2/20 - Loss: 1.1297 - Accuracy: 0.6923
Validation accuracy: 0.5968
Epoch 3/20 - Loss: 1.0895 - Accuracy: 0.7110
Validation accuracy: 0.8548
Epoch 4/20 - Loss: 0.5649 - Accuracy: 0.7483
Validation accuracy: 0.7742
Epoch 5/20 - Loss: 0.2582 - Accuracy: 0.7832
Validation accuracy: 0.8387
Epoch 6/20 - Loss: 0.1610 - Accuracy: 0.8089
Validation accuracy: 0.8710
Epoch 7/20 - Loss: 0.1333 - Accuracy: 0.8312
Validation accuracy: 0.8710
Epoch 8/20 - Loss: 0.2125 - Accuracy: 0.8453
Validation accuracy: 0.8548
Epoch 9/20 - Loss: 0.0219 - Accuracy: 0.8617
Validation accuracy: 0.9032
Epoch 10/20 - Loss: 0.0125 - Accuracy: 0.8755
Validation accuracy: 0.9032
Epoch 11/20 - Loss: 0.0087 - Accuracy: 0.8868
Validation accuracy: 0.9194
Epoch 12/20 - Loss: 0.0012 - Accuracy: 0.8963
Validation accuracy: 0.9032
Epoch 13/20 - Loss: 0.0009 - Accuracy: 0.9042
Validation accuracy: 0.9032
Epoch 14/20 - Loss: 0.0007 - Accuracy: 0.9111
Validation accuracy: 0.9032
Epoch 15/20 - Loss: 0.0006 - Accuracy: 0.9170
Validation accuracy: 0.9032
Epoch 16/20 - Loss: 0.0006 - Accuracy: 0.9222
Validation accuracy: 0.9032
Epoch 17/20 - Loss: 0.0006 - Accuracy: 0.9268
Validation accuracy: 0.9032
Epoch 18/20 - Loss: 0.0005 - Accuracy: 0.9308
Validation accuracy: 0.9032
Epoch 19/20 - Loss: 0.0005 - Accuracy: 0.9345
Validation accuracy: 0.9032
Epoch 20/20 - Loss: 0.0005 - Accuracy: 0.9378
Validation accuracy: 0.9032

Image processing - food classifier¶

Part I - Baseline Evaluation¶

Part II - Individual Algorithm Evaluation¶

Part III - Combination Search¶

Part IV - Model Fine-Tuning¶